GENOTYPE x ENVIRONMENT INTERACTION AND OPTIMAL NUMBER OF PROGENY TEST SITES FOR IMPROVING PINUS RADIATA IN NEW ZEALAND*

Transcription

1 GENOTYPE x ENVIRONMENT INTERACTION AND OPTIMAL NUMBER OF PROGENY TEST SITES FOR IMPROVING PINUS RADIATA IN NEW ZEALAND* S.D. CARSON Forest Research Institute, Private Bag 00, Rotorua, New Zealand (Received for publication 4 March 99) ABSTRACT A progeny test of 5 parents mated in a series of five, five-parent, disconnected diallels was established on sites chosen to represent all major site types for growing Pinus radiata D. Don in New Zealand. Statistical analysis of an assessment at age 9 years suggested that for diameter, the most important trait in the New Zealand tree improvement programme, genotype x environment interaction was important. However, genetic gains predicted for diameter for several regionalisation options, using multi-site index selection, suggested that regionalisation of seed orchards would increase the average genetic gain in diameter over all sites only slightly. Further, many fewer than the full sites were required for selection in order to capture essentially all of the predicted genetic gain for a national programme. Selection using the best site resulted in 90% of the maximum possible predicted gain, but the poorest site only %. Sites which were "very good" for selection had rapid growth, high phenotypic variance, and high repeatability of general combining ability (GCA) effects. Sites which were "very poor" for selection had slow growth, low variance, and low repeatability. Keywords: genotype x environment interaction; progeny testing; selection index; predicted genetic gains; specific combining ability; heritability; breeding strategy; Dothistroma pini; Cyclaneusma minus; Pinus radiata. INTRODUCTION Choice of the number of breeding and/or seed production regions within a tree improvement programme should depend on the trade-off between relative costs and benefits of regionalising. Breeding regions have often been initially defined on the basis of differing soil, climatic, or topographical conditions, with the assumption that different sets of tree genotypes will perform best in these varying environments. Differential performance of genetic entities grown under differing environments is an expression of "genotype x environment interaction" (GE). If GE is quantified in designed trials, by comparing the performance of common seedlots in differing regions, informed decisions can be made about the appropriate number of breeding regions. * Paper originally presented at IUFRO Conference, Olympia, Washington, USA, August 990. New Zealand Journal of Forestry Science (): -49 (99)

2 CarsonGenotype x environment interaction If GE is very pronounced, then for the same expenditure of resources a regionalised tree improvement programme would result in greater realised gains than would a non-regionalised programme. In situations of extremely high GE, it may be appropriate to operate separate breeding populations in different regions. Less extreme GE may be handled by producing different commercial seed orchard "breeds" for different regions, with each regional breed comprising a different set of genotypes selected from a single, national, breeding population. If GE is not large, a non-regionalised programme could yield greater gains for the same expenditure of resources, because of the increased selection intensities possible in a larger breeding population of select parents, and reduced progeny testing and implementation costs relative to those required for regionalised programmes. Costs of applied breeding programmes should always be evaluated against the potential gain from breeding. In discussing GE, Matheson & Raymond (984) proposed that the loss of potential genetic gain be used as a criterion to assess the level of practical importance of interactions. They went on to say, "Every breeder must have a level of loss of potential gain which is regarded as serious." Burdon (977) suggested "evaluating the expected genetic gains from the various possible options" for regionalising a breeding programme. These considerations guided this evaluation of GE. It has been suggested that selection for growth rate and adaptation to particular regions in New Zealand is warranted (Shelbourne et al. 987). This investigation set out to evaluate the importance of GE in improving Pinus radiata in New Zealand as reflected in a welldesigned trial planted on sites. The objective was to determine if GE is large enough to warrant a regionalised breeding and/or seed production programme with primary emphasis on the latter. GE was estimated for six important selection criteria on trees assessed at sites which represented the important site types for growing P. radiata in New Zealand. Predicted genetic gains from several different selection strategies were calculated and compared for diameter growth, the most important selection criterion in the New Zealand P. radiata improvement programme. The number of progeny tests required to represent a region can also be decided on the basis of the trade-off between relative costs and benefits of planting more tests. The optimum number of progeny tests required to attain a high genetic gain while minimising costs is an important consideration for any breeding programme. The high cost of progeny testing will usually place an upper limit on the number of test sites, while a lower limit will depend on the precision of breeding value estimates obtained at each site, risk of losing a test site, and the extent to which the site can be considered to represent the region. Equally important is to identify the particular test environments which give best resolution of genetic differences and, therefore, the most effective screening of progenies (Burdon 977). A second objective of this investigation was to examine the effect of varying the number and nature of progeny tests used for selection on genetic gain, in order to determine an optimum strategy for future progeny testing. EXPERIMENTAL METHODS Trials and Assessment A series of trials was planted at locations in 975 for the purpose of evaluating GE. Test seedlings were from 5 parents ("850" series) which were selected for improved growth and

3 4 New Zealand Journal of Forestry Science () form in the 950s from land-race stands in New Zealand. Parents were crossed in a series of five disconnected, modified, half-diallels (Griffing 956). Seed was sown in Rotorua and Rangiora nurseries in January 974 as five separate experiments for diallels one to five, and with four nursery replications of each diallel. Seedlings from the different nursery replicates were mixed during lifting, before allocation to field replications. Sites ranged from the northern part of the North Island to the southern part of the South Island (Fig. ), represented a wide range of rainfall and temperature conditions (Table ), and were selected to represent all major site "types" (Shelbourne et al. 987). S km 4 S U S FIG. -Eleven sites included in the genotype x environment interaction study Rotorua nursery supplied plant material for sites at Maramarua, Kaingaroa (Cpt 7 and 905), Awahohonu, Golden Downs, and Mawhera. Rangiora nursery supplied Woodhill, Ruatoria, Eyrewell, Berwick, and Taringatura. Six block replicates (reps) (each 0.7 ha) were planted at each site in 975 at 5 x 5 m spacing, with each block replicate divided into five sub-blocks. Within each block replicate, each of the five diallels was randomly assigned to one of the sub-blocks. Five trees from each full-sib cross per diallel were randomised as single-tree plots within that diallel's sub-block. There were, therefore, 0 trees per full-sib cross and 0 trees per parent represented at each trial site. Six traits were assessed at each site 9 years from planting, including: diameter in centimetres at.4 m height on the uphill side of the tree (dbh),

4 TABLE Site latitude, rainfall, temperature, and diameter mean, repeatability, phenotypic variance, genetic (above diagonal) and phenotypic correlations of breeding values Latitude Mean annual rainfall (mm)* Mean annual temp. ( C)t Mean Repeatability of half-sib family effects Phenotypic variance of half-sib family effects Site: Mean phenotypic correlation with all sites Relative importance ofsca Woodhill 5 00' % Maramarua 7 5' % Ruatoria 7 54' % * From New Zealand Meteorological Service (98). f From Esk Station. $ From Totara Flat Station. From Winton Station. 4 Kaingaroa Cpt 7 8 0' % 5 Kaingaroa Cpt ' % 6 Awahohonu 9 0' 756t.9t % 7 Golden Downs 4 4' % 8 Mawhera 4 7' 89$.$ % 9 Eyrewell 4 5' % 0 Berwick 45 58' % Taringatura 45 59' % All sites

5 6 New Zealand Journal of Forestry Science () straightness using a -9 subjective score where = crooked, 9 = very straight, branch habit using a -9 subjective score where = extremely uninodal, 9 = extremely multinodal, malformation using a -9 subjective score where 9 = no malformation, 8 = ramicorn, 7= ramicorns or ramicorn intermediate in diameter between normal ramicorn and a fork, 6 = or more ramicorns or severe plus one average ramicorn, 5 = basket whorl with stem not deflected, 4 = basket whorl with stem deflection up to half diameter, = basket whorl with stem deflection greater than diameter, = fork, = more than one fork, Dothistroma needle blight using a -9 scale reflecting the percentage of crown length infected where = 0% infected, and 0 = 00% infected (two sites only), needle retention using a -6 scale reflecting the number of years' needles remaining on a branch at mid-crown (or above the level of Dothistroma infection) where = about half of first-year needles retained, 6 = all first-, second-, and third-year needles retained (0 sites only). Infection by Dothistroma pini Hulbary reduces growth in proportion to disease severity (van der Pas 98). Needle retention is related to infection by Cyclaneusma minus (Butin) DiCosmo, Peredo & Minter (Gadgil 984), and its effect on growth has also been demonstrated (van der Pas, Slater-Hayes, Gadgil & Bulman 984). Evaluation of Genotype x Environment Interaction Arithmetic full-sib family (cross) means were calculated for each sub-block for each trait at each site. Separate analyses of variance were made for each diallel at each site using program DI ALL (Schaffer & Usanis 969) and following the linear model: Y ijk = (I + b k + gj + gj + Sjj + e ijk where Yp is the mean score of the progeny of the i th female and j th male in the k th block replicate; (I is the mean score for the diallel; b k is the effect of the k th replicate; g\ is the general combining ability (GCA) effect of the i th female; gj is the GCA effect of the j th male: sy is the specific combining ability (SCA) of the cross between the i th female and j th male; and ejj k represents an error effect (Wilcox 98). Griffing's Method 4 (Model I), which assumes GCA and SCA effects are fixed, was followed for the estimation and testing of GCA and SCA effects (Griffing 956) in each diallel. Parental breeding values at each site, defined as BVj = gi + X, were estimated as outlined by Snyder (975). Narrow-sense heritability (h ), repeatability of half-sib family effects (h hs), and phenotypic variance of individuals (G P) and of half-sib family effects (o P ) were estimated for each trait at each site as follows: h = ^ A = 4 <> GCA ^, and GCA T + w CJVA SCA + &

6 CarsonGenotype x environment interaction 7 h hs o n GCA GCA p hs <> GCA + SCA / + <? P / 8 + o w / 8 H where <J GCA = general combining ability variance, a p = phenotypic variance of individuals, a - = phenotypic variance of half-sib family effects, G SCA = specific combining ability variance, o = variance due to interaction between full-sib families and block replicates, G W = within sub-class variance, and H = harmonic mean of number of trees within each subclass (which ranged from.4 to 4.). For the purpose of estimating GCA and SC A variance components, Griffing' s Method 4 (Model II) was followed (Griffing 956; Wilcox 98). This entails no change to the method of calculating mean squares in the analysis of variance of each diallel, though their expected values are different to allow for the assumption that GCA and SCA effects are now random. Since the structures of the five diallels were compatible, sums of squares from the individual diallels were pooled to give an overall analysis, and estimates of GCA and SCA variance components based on all 50 full-sib crosses and all 5 parents. An analysis of variance for each trait over all sites was carried out using PROC GLM in SAS (SAS Institute Inc. 985) with the following model: Y ijk i m = i + Sj + b(s) y + D k + SD ik + Db(S) ijk + C(D) kl + SC(D) ikl + Cb(SD) ijkl +e ijklm where Yjj k i m is the score of the m th tree from the th full-sib family cross in the k th diallel in the j t h block at the i th site; (I is the mean score overall diallels and sites; Sj is the effect of the i th site; b(s)y is the effect of the j t h block at the i th site; D k is the effect of the k th diallel; SD ik is the effect of the k th diallel at the i th site; Db(S) ijk is the effect of the k th diallel in the j t h block at the i th site; C(D) kl is the effect of the th cross within the k th diallel; SC(D) ikl is the effect of the th cross within the k th diallel at the i th site; Cb(SD)jj kl is the effect of the th cross within the k th diallel in the j l h block at the i th site; and eyki m represents an error effect. Variance components were calculated by equating estimated mean squares to their expectations as outlined by M.D.Wilcox (unpubl. data). Sums of squares for the first eight terms of the model were calculated using sub-block means. The sum of squares for the error term was calculated in separate analyses by treating all sub-blocks as separate treatments. The model provided direct tests and estimates of the variance of crosses within diallels (o C(D)) and the interaction variance of crosses x sites within diallels (o SC(D)). The ratio of the interaction variance of crosses x sites within diallels to the variance of crosses within diallels was calculated for each trait, as suggested by Shelbourne (97). For the purpose of estimating GCA and SCA variance components over all sites, each diallel was analysed over all locations using programme DIALL and following the linear model: Yijkm = ^ + gi + gj + sy + S m + b(s) km + gs im + gs jm + ss ijm + e ijkm where l, gj, gj, and sy are defined as before; Yy k i is the mean score of the i th female and the j t h male in the k th block at the m th site; S m is the effect of the m th site; b(s) km is the effect of

7 8 New Zealand Journal of Forestry Science () the k th block at the m th site; gsj m and gsj m are interaction effects involving GCA and sites; and ssjj m is the interaction effect between the SCA of the ij th cross and the m th site (Wilcox unpubl. data). Griffing's Method 4 (Model II) was again followed (Griffing 956). Sums of squares from the individual diallels were again pooled to give an overall analysis and estimates of GCA, SCA, GCA x site interaction, and SCA x site interaction variance components based on all 50 full-sib families and all 5 parents over all sites. The ratio of GCA x site interaction variance to GCA variance was calculated (Shelbourne 97). Narrow sense heritability (h ), repeatability of half-sib family effects (h hs), and phenotypic variance of individuals (a P) and of half-sib family effects (o p. ) were estimated for each trait over all sites as follows: u 4<J GCA 4<J GCA o, a = ^~ = j ~ ' P G GCA + G SCA + GCAxS + SCAxS + p + w h\ c = GCA hs v GCA rr_ r? a a.rf /^, ^ /n4.rr / Q n 4- r? / Q n 4- ro- Phs GCA + a sca / + a GCAXS / n + a SCAxS / n + a / 8n + a w / 8nH where CJ GCA = general combining ability variance estimated from all sites, G P = phenotypic variance of individuals, "Phs a SCA a u n GCAxS a u SCAxS P a w H phenotypic variance of half-sib family effects estimated from all sites, specific combining ability variance estimated from all sites, general combining ability by site interaction variance, number of sites, specific combining ability by site interaction variance, variance due to interaction between full-sib families and block replicates estimated from all sites, within sub-class variance estimated from all sites, and harmonic mean of number of trees within each sub-class over all sites. Estimates of the relative importance of specific combining ability (R) (Baker 978) were calculated both for each trait over all sites and for each trait at each site as R = SCA / ( ^GCA + SCA) >< 00% ' Genetic correlations of diameters at different sites were approximated as follows (Burdon 977): h A xh B where r G = genetic correlations of parental GCAs or GCA correlations for diameter from site A with site B (an approximation to additive genetic correlations), r p = correlation of parent breeding values from site A with site B, h A = square-root of repeatability of half-sib family effects for site A, h B = square-root of repeatability of half-sib family effects for site B. This method can lead to genetic correlations greater than one, especially with exceptionally low repeatability estimates.

8 CarsonGenotype x environment interaction 9 Predicted Gains from Different Selection Strategies Predicted genetic gains from selection for increased dbh were calculated for several regionalisation options using genetic selection indices (Hazel 94) adapting the method of Burdon (979) whereby the expression of a trait at each site is handled as a separate index trait. The index was of the form I = bj Sj + b S bjj Sjj where Sj to Sj j are parent breeding values for dbh at sites to, respectively, and b x to b x Y are the corresponding index coefficients. Programme RESI (Cotterill & Jackson 98), which is based on equations presented by James (968), was used to calculate all indices and predicted gains. A single selection intensity was chosen to approximate the operational selection of best parents in the New Zealand P. radiata improvement programme for use in control-pollinated clonal seed orchards, that is, a choice of 0 parents out of a breeding population of about 50. In order to simulate the performance of a single, national, improved seed product, predicted gains in dbh at each site were compared for selection of one set of parents for New Zealand, using four alternative sets of economic weights. Economic weights for each site were either () one, () the reciprocal of the mean diameter of all trees at that site (in order to achieve the same percentage gain on all sites), () the proportional area of P. radiata plantations represented by the site as defined using data presented by Stewart & Burrows (986) and I. R. Hunter (pers, comm.), or (4) reciprocal of the mean diameter of trees at the site multiplied by the proportional area of plantations represented by the site. The effect of selecting different sets of seed orchard parents to produce improved seed for different forest regions was then modelled. Nine different regionalisation strategies were devised, using hypotheses of regional structure based on inspection of correlations of parental GCAs among sites for dbh, rank changes in breeding values over sites, and local knowledge about how P. radiata trees grow on New Zealand sites. For all of the nine regionalisation strategies, each of the sites in the trial was allocated to hypothetical regions. For each of the nine strategies selection indices were used to calculate predicted gains for each hypothetical region. Information from all sites was used in each index with economic weights of zero assigned to sites not in the designated region. This approach simulates selection using all information available, whether it be from within a region or from outside the region. Best Sites for Selection For either a national or a regionalised seed production programme, it is necessary to determine both the appropriate number and the location of progeny trials needed to achieve maximum genetic gains. To simulate the selection of one set of parents for countrywide use, predicted gains from every possible combination of selection sites were calculated using selection indices with Binet restrictions (Binet 965; Cotterill & Jackson 98). The Binet restriction placed on a site forces the coefficient in the index for that site (the b value) to be zero, allowing the predicted gain on that site to be calculated without using information from that site for selection (as would result from a national seed production programme that relies on progeny test data from a limited number of sites).

9 40 New Zealand Journal of Forestry Science () Selection indices were calculated using RESI with the Binet restriction to obtain average predicted gains over all sites from all possible combinations of one site, two sites, three sites, etc., up to 0 sites. The stand error of prediction was calculated for the average predicted gain over all sites from the index using data from sites (White & Hodge 989). For each possible number of progeny test sites (from to ), the mean overall gain and standard deviation for all possible site combinations were calculated and plotted. The number of times each site occurred in the four "best" combinations of sites and the four "worst" combinations of sites were counted for sets of one, two, three, four, and five sites used for selection. "Best" and "worst" were determined by having the highest and lowest overall gains, respectively. RESULTS Evaluation of Genotype x Environment Interaction Results from analyses of variance suggested that GE interaction might be important for only two of the six traits assessed, dbh and needle retention. Although there were significant crosses x site interactions for all traits except Dothistroma infection (Table ), the F statistics were not much greater than one (all but one were less than.0). These F statistics were tested with high degrees of freedom in the denominator (404 for the analyses with data from sites), with the result that even very small F statistics were significant in a statistical sense. For example, an F of. was highly significant for malformation. Further, there were TABLE -Genetic parameters fornix traits of Pinus radiata assessed on all sites at age 9 Dbh (cm) Number of sites: Crosses (diallels) a C(D) Straightness (-9) Branch habit (-9) Malformation (-9) Needle retention (-6) 0* Dothistroma (-0) t 48 ' 9 ' ' 56 F$ 7.7** 8.5**.**.9** 5.84**.7** Sites x Crosses (diallels) c SC(D) F.79**.5**.46**.**.**.4 NS o SC(D)/G C(D) Sites x GCA (diallels) a GCAxs/ a GCA - 70 * 6 ' 4 0J9 * 86 ' 57 Relative importance of SCAD % 8% % 0% 4% 8% Narrow-sense heritability and repeatability of half-sib family effects h h hs ** Statistically significant at p < 0.0. * Berwick not included. t Kaingaroa Cpt 905 and Mawhera. $ 45/450 d.f. for sites. 450/404 d.f. for sites. <> SCA / (^GCA + SCA) X 00% '

10 CarsonGenotype x environment interaction 4 heterogeneous variances among sites, which could well have been the source of the significant interaction, rather than rank changes. The ratio of the interaction variance of crosses x sites to the variance among crosses was less than 0.5 for straightness and branch habit, suggesting less important GE (Shelbourne 97). GCA variance was much more important than SCA variance for all traits (Table ). The ratio of GCA x site interaction variance to GCA variance was less than the ratio of the interaction variance of crosses x sites to the variance of crosses for every trait except Dothistroma infection, which was only slightly more than 0.5. Dothistroma infection was similar for the two sites on which it was measured, with an average of 0% and 4% of the needles infected and a similar range of cross means (Carson 989). Cross means for malformation differed significantly on only two sites, and heritability of this trait over all sites was particularly low (0.0). Taken together, these results suggest that GE is not important in P. radiata in New Zealand for Dothistroma infection, straightness, branch habit, or malformation. For dbh and needle retention, the F statistics for the crosses x site interaction and the ratio of the interaction variance to the variance among crosses were largest for the six traits (Table ). Both dbh and needle retention are used as selection criteria for P. radiata in New Zealand, but dbh is always given greater emphasis than other traits, while needle retention is of indirect importance through an influence on later diameter growth. Extensive analysis of the practical importance of GE was, therefore, carried out for dbh. The sites of this trial produced a wide range of different growth rates for P. radiata families (Table ). Mean dbh ranged from 9.8 cm at Berwick to 5.7 cm at Kaingaroa Forest, Cpt 7. The repeatability of half-sib family effects was highly correlated with mean dbh (r=0.9, p<0.00). Phenotypic variance for dbh also ranged widely, from 0 to 58, and was largest on the sites with better growth. Diallels were not significantly different on any site. Over all sites, the estimate of the repeatability of half-sib family effects for dbh was 0.75, which is high enough to indicate that selection of parents using breeding values calculated from this trial would be successful in achieving moderate to high average gains in dbh for the country as a whole. Selection using individual-site information for nationwide gains, however, would achieve varying success, since the estimated repeatability of half-sib family effects varied from a very low 0.0 to 0.8 for individual sites (Table ). No clear regional pattern of site combinations was apparent from inspection of the pairwise correlations of breeding values (Table ). There was no indication that nursery effects carried over in the between-site correlations. Phenotypic correlations between sites ranged from -0. to 0.8. The mean of the correlations of any one site with other sites was higher for some sites than others, ranging from 0. to 0.55 (Table ), with each site exhibiting a fairly wide range of correlations with the other sites. Inspection of rank changes of parent breeding values for dbh (details not shown) also failed to reveal a clear pattern favouring a regionalised strategy. There were large changes in ranks of some parents from site to site. For example, parent 0 ranked first at one site and twentieth at another, parent 55 ranked first or second at six sites yet ranked eighteenth at two sites, and parent 0 ranked first at two sites, yet last at another. Despite these large rank changes, however, no clear-cut site-to-site pattern emerged. There was no evident tendency for groups of parents to change rank together, but rather each parent tended to vary differently from site to site.

11 4 New Zealand Journal of Forestry Science () Predicted Gains from Different Selection Strategies Predicted gain in dbh was affected very little by varying economic weights for different sites in the genetic selection index. Averaged gain over all sites for the four different sets of economic weights ranged from.0 cm (or.%) to.5 cm (or.5%). Differences in predicted gain for the four different sets of economic weights on a single site were nevermore than 0.6 cm, and were almost always in the range cm. Of the four types of economic weighting compared, the reciprocal of the site mean diameter was used for the remainder of the analyses. Predicted overall gain in dbh did vary among the nine hypothesised regionalisation schemes (Table ). The differences were, however, not very great even for the extreme case of comparing selection of one set of parents for the whole country (. cm or.% gain) with selection of differing sets of parents for regions (.7 cm or 4.4% gain) (columns &, Table ). Differences in predicted gains from national versus regionalisation strategies involving only two or three regions were even smaller. The maximum predicted gain from selection strategies using either two or three regions was.4 cm or.8%. Both of these schemes involved at least one region represented by only one site, raising the question of whether within-region variation among sites was adequately represented and, therefore, whether predicted gains could be realised. The maximum gain from strategies using regions with two or more sites was. cm or.8%, which is only marginally greater than gains for a national strategy. Best Sites for Selection Both the number of progeny-test sites used for selection and the choice of specific sites influenced predicted genetic gains (Fig. ). Far fewer than the full sites were required for selection to capture essentially all of the predicted genetic gain. Some sets of two, three, or four sites yielded predicted gains very close (above 95%) to the maximum predicted gain obtained when using all sites for selection. Even use of a random choice of five sites would result in a predicted gain of more than 80% of the maximum. The standard error of the prediction for gain using all sites for selection was % of the predicted gain. 00 ^M~"~~~\7-~ """* ^4r- "" average predicted gain standard deviation maximum and minimum Number of sites \ I I I FIG. -Average predicted gain over sites when different sites and different numbers of sites are used for selection of parents for improvement of diameter

12 TABLE -Gain in dbh (cm) predicted for nine different regionalisation schemes. Indices were constructed to represent selection of 0 out of 50 Pinus radiata parents for control-pollinated orchards. Regionalisation scheme Region Gain Region Gain Region Gain Region Gain Region Gain Region Gain Region Gain Region Gain Region Gain Woodhill Maramarua Ruatoria Kaingaroa, Cpt 7 Kaingaroa, Cpt 905 Awahohonu Golden Downs Mawhera Eyrewell Berwick Taringatura [ I..7 I..9 I..8 I.7.0 L.8.9 I.0.6 I.0.6 I.6. I No. regions in regionalisation scheme Av. gain over all sites

13 44 New Zealand Journal of Forestry Science () Some sites were consistently good for selection, while others were consistently poor. When Cpt 7 at Kaingaroa Forest was used alone for selection, predicted gains were 90% of the maximum possible predicted gain. Conversely, when Maramarua was used as the only selection site, gains were only % of the maximum possible. For selection indices using information from one to five sites, individual sites could be classified either as frequently being in sets of sites which had the highest predicted gains, or frequently being in sets of sites with the lowest predicted gains (Table 4). The sites which would be "very good" for selection (Woodhill, both sites within Kaingaroa Forest, and Awahohonu) had high phenotypic variances and high repeatabilities of family means, and the sites which would be "very poor" (Maramarua, Eyrewell, Berwick, and Taringatura) had low variances and low repeatabilities. Average correlations with other sites tended to be higher for the "very good" sites than for the "very poor" sites (Table ). TABLE 4Number of times a site occurred in the best and worst sets of sites using to 5 sites for selection Site Woodhill Maramarua Ruatoria Kaingaroa, Cpt 7 Kaingaroa, Cpt 905 Awahohonu Golden Downs Mawhera Eyrewell Berwick Taringatura 4 best sets (highest predicted gain) worst sets (lowest predicted gain) Worth as selection site Very good Very poor Poor Very good Very good Very good Moderate Poor Very poor Very poor Very poor DISCUSSION AND CONCLUSIONS Importance of Genotype x Environment Interaction The estimate of repeatability of half-sib family effects at the Berwick site was exceptionally low (0.0). A trial of 04 polycrossed families from the same selection series ("850" series) (CJ.A.Shelbourne unpubl. data) was planted adjacent to the Berwick site and had an estimate of family mean repeatability of 0.6. The Berwick site had many missing and damaged trees as a result of severe weather conditions and was partially thinned in error, so it was generally of very poor quality as a progeny trial. The estimate of the SCA variance was very high as compared with that of GCA variance (Table ), probably a result of poor trial conditions rather than a real difference in importance of SCA at this site. The same applies to some extent to the Eyrewell site (estimated repeatability of half-sib family effects 0.5), which also had a polycross trial planted adjacent (estimated repeatability of family means 0.74). Phenotypic correlations of breeding values for these two sites with the other sites, however, were in a range similar to the other sites (Table ) and to those calculated in other studies (Johnson & Burdon 990; Shelboume unpubl. data; Shelboume & Low 980), and so subsequent analyses and interpretations are thought to be valid.

14 CarsonGenotype x environment interaction 45 The size of the GE interaction in P. radiata in New Zealand appears to be too small to warrant regionalised breeding populations. A tree improvement programme with regionalised breeding populations would by necessity have smaller breeding populations in each region, lower selection intensities, and, therefore, lower genetic gains than a national programme of the same size. However, even a regionalised selection programme (with a national breeding population) for seed production would be more expensive to operate than a non-regionalised programme for several reasons. Seed orchard land requirements would be somewhat greater, and operating costs proportionally higher. Additional records and extension efforts would be required to ensure that the appropriate regional breeds were planted in the specified regions. Further, substantial additional costs arise from the progeny-testing phase. In the analysis presented here, progeny test size and selection intensities were assumed to be the same for each region in a regionalised programme as for a non-regionalised programme. In practice, regionalisation of seed production with fixed costs for progeny testing would reduce selection intensities and/or precision applied within each region, which in turn reduces the magnitude of predicted gain. In order to achieve a similar selection intensity to that for a nonregionalised programme, an overall bigger and more expensive progeny-testing or reselection programme would be required, since each potential seed orchard parent must be represented in a number of progeny tests in every region. The additional costs of progeny testing would need to be offset by substantial increases in genetic gain. Given the data presented here, regionalisation of commercial seed production for the New Zealand P. radiata programme for the purpose of obtaining increased diameter growth would not be economically sound. A fully regionalised seed production programme, that is, one in which separate selections of seed orchard parents would be made for each of the sites in the trial, did increase predictions of genetic gain across New Zealand from.% to 4.4% of the mean dbh at age 9. Although this represents a fairly substantial increase in gain, it is probably not realisable, since genetic gains calculated from regions represented by one site cannot be considered to broadly apply to a whole region. The assumption that the site exactly represents the region (that is, that the phenotypic and genetic correlations with other sites within the region are one) is very unlikely to be met, especially in heterogeneous regions such as, for example, the phosphate-retentive soils in the Northland region. Even on sites as similar as are those on the volcanic pumice soils in the central North Island region (represented here by the two sites in Kaingaroa Forest and Awahohonu), phenotypic and genetic correlations are less than unity. The gain estimate for "regions" represented by more than one site is probably a more realistic one. The small increase in predicted gain (0. cm, or 0.6% of mean dbh at age 9) probably does not outweigh the cost of regionalisation, and resources spent in other areas are likely to yield greater returns for the expenditure. The small increase in predicted genetic gain with regionalisation of this type suggests that for the sites represented the amount of "within region" GE is comparable with the overall New Zealand-wide GE for P. radiata. Other studies of GE for diameter growth in P. radiata have reached a similar conclusion. Shelbourne & Low (980) found that expected gains for three sites were only slightly less than those from selecting families specifically for each site. Another New Zealand study compared two sites from the phosphate-retentive clay soils in the Northland region (which includes the Maramarua site, the site this study suggested was the most different from the

15 46 New Zealand Journal of Forestry Science () others) with two sites on volcanic pumice soils in the central North Island region (Johnson & Burdon 990). Although strong genetic correlations were observed between sites within each of the two regions and relatively weak correlations between sites in different regions, predicted genetic gains from regionalisation were still not sufficient for the authors to advocate regionalised seed production. Average gains over all four sites in the study were greater, however, when sites from both regions were used for selection. Although the data from the trial reported here suggest that the Maramarua site would be very poor for selection with its low repeatability and low phenotypic variance, this site had less than ideal progeny test conditions because of severe weed competition. It is possible that better quality progeny tests in the Northland clay region could raise the overall genetic gains in that region. However, this is likely to have only a small impact on raising the average genetic gain over all New Zealand, because the Northland clay region is relatively small and was the only region represented in this trial which appeared to behave in any way anomalously with respect to family rankings. Results from studies of P. radiata in Australia suggested that GE may be of greater importance in Australia than in New Zealand (ratios of GE variance to family variance were higher) (Matheson & Raymond 984), but did not include an analysis of predicted gains. Even so, the authors did not advocate a regionalised breeding programme and suggested instead that the most interactive parents be discarded. Discarding the most interactive parents, however, does not appear to be an ideal approach, because identification of the interactive parents requires a larger progeny-test programme for very little, if any, additional gain. In addition, the best parents overall can be the most interactive, as was apparent in the performance of parents reported here. The index approach quantified the effects of GE in terms which were easily evaluated. Comparisons of predicted gains in diameter from regionalised versus non-regionalised seed production strategies allowed evaluation of the practical importance of GE. Even with some of the estimated genetic correlations in excess of unity, the majority of the indices did not have counter-intuitive negative weights. Using untransformed site variances allowed comparison of the magnitude of expected gains per site for different regionalisation strategies. Comparison of predicted genetic gains puts the effect of GE into realistic terms, rather than just abstract statistics. Results for several other traits assessed in this trial also failed to support regionalisation due to GE. Statistics for Dothistroma infection, stem straightness, branch habit, and stem malformation indicated that there was substantially less GE in these traits than for stem diameter, suggesting that a regionalised programme would not result in much increased gain in these traits. GE has not been reported for needle retention before, but growth rate has been found to show more GE than tree-form traits in several New Zealand studies (Johnson & Burdon 990; Shelbourne unpubl. data; Wilcox unpubl. data), but about the same low level of GE in another study (Shelbourne & Low 980). GE for Dothistroma resistance was also small in another study involving different parents planted on different sites (Carson 989). Needle retention did appear from variance-component estimates to have as great a GE relative to GCA effects as dbh. Pathogenicity tests have shown that needle loss is associated with infection by Cyclaneusma minus (Gadgil 984). Needle retention, however, appears to involve a complex set of variables rather than simply being closely related to the size of the fungus population and the suitability of climatic conditions for spore production and

16 CarsonGenotype x environment interaction 47 dissemination (as with Dothistroma needle blight). On four pine species in the Pacific Northwest (not including P. radiata) C. minus has been identified as an endophytic fungus, that is, one which is commonly isolated from green needles (Carroll & Carroll 978). Infection first occurs in P. radiata in New Zealand when needles are about 6 months old (Gadgil 984). Needle-cast may occur when needles are about a year old or not until they are 0- months old. It is possible that when symptoms appear soon after infection, it is because the tree is under stress of some sort, rather than simply because the fungus has been able to infect a susceptible genotype. This stress might be related to different factors on different sites. The New Zealand breeding programme should not at this point be regionalised for needle retention. Predicted genetic gains were not calculated for this trait, but would probably present a similar story to dbh. In addition, it may be that needle loss is related to different stresses on trees in different regions (for example, drought, wind, poor soil, cold climate). Resistance to each of these stresses can probably be identified with further research, but each is probably under different genetic control, that is, independently inherited. If this is the case, regional selection becomes more difficult and expensive because data from one region cannot be used to select trees in another region. Much more should be known about the biology of needle retention before a regionalised selection programme is considered for this trait. Other Reasons for a Regionalised Programme Results from this investigation suggest that regionalisation as a response to GE which is found in New Zealand would not be warranted. However, there may be other reasons for pursuing regionalised breeding and/or selection programmes. For example, the relatively higher cost of regionalisation could become warranted by higher gains if different sets of selection criteria are appropriate in different regions. One such example arises with internode length, which is expressed differently on some New Zealand sites. Average internode length varies as much with site as with genotype, and variance among families depends strongly on site (Carson & Inglis 988). It might be desirable to include internode length as a selection criterion on the pumice plateau where there are larger differences among families. However, in Southland all trees tend to have long internodes and in Northland all trees tend to have short internodes, and so genetic gains in internode length would be much smaller. It might, therefore, be desirable to distinguish between long internode and short internode breeds for use on the pumice plateau, while concentrating on use of a single growth and form breed in Southland and Northland. Another example concerns the relative weighting of growth and tree-form traits in regions where tree form tends to be generally good, such as in the coastal sands (represented by the Woodhill site), compared with regions where tree form is highly variable and varies beyond the threshold of desirability. Regionalisation of selection criteria might be warranted by higher gains if there are regional demands for different end products. If wood from a given region is used primarily in a pulp mill, selection criteria could be designed around requirements for that mill. In contrast, wood intended primarily for clearwood export markets might suggest a different set of selection criteria.

17 48 New Zealand Journal of Forestry Science () Regionalisation of selection criteria might also give significantly higher gains than a nonregionalised programme if there are specific problems on some sites. For example, Dothistroma needle blight is severe in roughly 0% of New Zealand's P. radiata forest (van der Pas, Bulman & Horgan 984). In these areas major emphasis on selection for Dothistroma resistance in both seed production and breeding populations is desirable to maximise yield (Carson et al. 99). In the other forests there is little impact from the disease, presumably because climatic conditions which have been identified as suboptimal for survival and reproduction of the causal agent (the fungus Dothistroma pini) (Gadgil 977) prevail in these areas. Because gains in growth (the most important selection criterion) in the absence of the disease are reduced when resistance is included as an additional selection criterion (Carson 989), it is desirable to place little emphasis on resistance to Dothistroma needle blight as a selection criterion both in seed production and breeding populations for sites where the disease has little or no impact. Choice of Selection Sites Comparisons of predicted gains for selection across the entire country, calculated from different numbers of sites and different sets of sites in genetic selection indices, suggest that for P. radiata in New Zealand sites can be classified as either "good" or "poor" selection sites. A few well-chosen selection sites could yield as much gain as many poorly chosen sites. Good selection sites were characterised by rapid growth, high repeatability of GCA effects, high phenotypic variance, and on average the highest correlations with other sites, while poor selection sites had low repeatability, small phenotypic variance, and the lowest correlations with other sites. Although this now appears to be intuitively obvious, before this analysis was done the poor sites often tended to beidentified as those which indicated a need for regional seed production programmes, and they were therefore seen to be important progeny test sites. ACKNOWLEDGMENTS CJ.A.Shelbourne had the foresight to design and implement the trial. G.Vincent, A.Firth, and C.Low were responsible for the planting, assessment, data management and analyses, respectively, with assistance from many others in the Genetics and Tree Improvement group. Unpublished work by M.D.Wilcox, which outlined expectations of mean squares and methods of diallel analysis, was used as a guide throughout this study. The author wishes to thank R.D.Burdon and G.R.Johnson for early discussions of the method, M.Carson, R.D.Burdon, and CJ.A.Shelbourne for thorough review of the manuscript, and P.Cotterill and J. King for encouragement with the project. REFERENCES BAKER, R.J. 978: Issues in diallel analysis. Crop Science 8: 5-6. BINET, F.E. 965: On the construction of an index for indirect selection. Biometrics : 9-9. BURDON, R.D. 977: Genetic correlation as a concept for studying genotype-environment interaction in forest tree breeding. Silvae Genetica 6: : Generalisation of multi-trait selection indices using information from several sites. New Zealand Journal of Forestry Sciience 9: 45-. CARROLL, G.C.; CARROLL, F.E. 978: Studies on the incidence of coniferous needle endophytes in the Pacific Northwest. Canadian Journal of Botany 56: CARSON, M.J.; INGLIS, CS. 988: Genotype and location effects on internode length of Pinus radiata in New Zealand. New Zealand Journal of Forestry Science 8:

18 CarsonGenotype x environment interaction 49 CARSON, S.D. 989: Selecting Pinus radiata for resistance to Dothistroma needle blight. New Zealand Journal of Forestry Science 9: -. CARSON, S.D.; DICK, A.M.P.; WEST, G.G. 99: Benefits of the Dothistroma-resistant breed of radiata pine. Pp. 5-6 in Whyte, A.G.D. (Ed.) "New Directions in Forestry, the Costs and Benefits of Change", Proceedings Australia New Zealand Institute of Foresters Conference, Sept-Oct 99. COTTERILL, P.; JACKSON, N. 98: ShortNote: Index selection with restrictions in tree breeding. Silvae Genetica 0: GADGIL, P.D. 977: Duration of leaf wetness periods and infection of Pinus radiata by Dothistroma pini. New Zealand Journal of Forestry Science 7: : Cyclaneusma (Naemacyclus) needle-cast of Pinus radiata in New Zealand.. Biology of Cyclaneusma minus. New Zealand Journal of Forestry Science 4: GRIFFING, B. 956: Concept of general and specific combining ability in relation to diallel crossing systems. Australian Journal of Biological Sciences 9: HAZEL, L.N. 94: The genetic basis for constructing selection indexes. Genetics 8: JAMES, J.W. 968: Index selection with restrictions. Biometrics 4: JOHNSON, G.R.; BURDON, R.D. 990: Family-site interaction in Pinus radiata: Implications for, progeny testing strategy and regionalised breeding in New Zealand. Silvae Genetica 9: MATHESON, A.C.; RAYMOND, CA. 984: The impact of genotype x environment interactions on Australian Pinus radiata breeding programs. Australian Forest Research 4: -5. NEWZEALAND METEOROLOGICAL SERVICE 98: Summaries of climatological observations to 980. New Zealand Meteorological Service, Wellington. SAS INSTITUTE INC. 985: "SAS User's Guide: Statistics". Version 5 Edition. SAS Institute Inc. Box 8000, Cary, NC SHAFFER, H.E.; USANIS, R.A. 969: General least squares analysis of diallel experiments. North Carolina State University, Genetics Department Research Report Number 7. 6 p. SHELBOURNE, C.J.A. 97: Genotype-environment interaction: Its study and its implications in forest tree improvement. Proceedings of IUFRO Genetics-SABARAO joint Symposium, Tokyo B-l(I): -8. SHELBOURNE, C.J.A.; LOW, CB. 980: Multi-trait index selection and associated genetic gains of Pinus radiata progenies at five sites. New Zealand Journal of Forestry Science 0: SHELBOURNE, C.J.A.; BURDON, R.D.; CARSON, S.D.; FIRTH, A.; VINCENT, G. 987: "Development Plan for Radiata Pine Breeding". Forest Research Institute, Rotorua, New Zealand. 4 p. SNYDER, E.B. 975: Combining-ability determinations for incomplete mating designs. USDA Forest Service, Southern Forest Experiment Station, General Technical Report SO-9. STEWART, H.; BURROWS, G. 986: A national exotic forest description system. New Zealand Forestry Council, Working paper No...95 p. van der PAS, J.B. 98: Reduced early growth rates of Pinus radiata caused by Dothistroma pini. New Zealand Journal of Forestry Science 77:0-0. van der PAS, J.B.; BULMAN, L.; HORGAN, G.P. 984: Disease control by aerial spraying of Dothistroma pini in tended stands of Pinus radiata in New Zealand. New Zealand Journal of Forestry Science 4: -40. van der PAS, J.B.; SLATER-HAYES, J.D.; GADGIL, P.D.; and BULMAN, L. 984: Cyclaneusma (Naemacyclus) needle-cast of Pinus radiata in New Zealand. Reduction in growth of the host and its economic implications. New Zealand Journal of Forestry Science 4: WHITE, T.L.; HODGE, G.R. 989: "Predicting Breeding Values with Applications in Forest Tree Improvement". Kluwer Academic Publishers, Dordrecht, The Netherlands. 67 p. WILCOX, M.D. 98: Genetic variation and inheritance of resistance to Dothistroma needle blight in Pinus radiata. New Zealand Journal of Forestry Science : 45.